With the rapid expansion of the electric vehicle market, the number of power batteries reaching their end-of-life stage has grown dramatically. It is estimated that by 2025 the retired battery volume will exceed 820 thousand tons globally, and by 2028 the annual retired amount will surpass 2.6 million tons. These retired batteries are not merely waste products; they contain valuable metals such as cobalt, nickel, manganese, and lithium. If they are not properly treated, severe environmental pollution and a significant waste of resources will be caused. On the other hand, the recycling market is still at an early stage, and the cost of dismantling, transporting, and evaluating retired power batteries remains high. One promising solution is to estimate the recovery value on-site before any complex dismantling operation, using only surface-level information that can be observed or accessed conveniently. However, the relationship between these surface data and the final recycling price is highly nonlinear, noisy, and affected by many hidden factors. The mechanism that connects the externally visible features of a retired battery to its monetary value is still poorly understood. As a result, the accuracy of existing recycling price prediction methods is often unsatisfactory.
In this article, we propose a systematic method for predicting the recycling price of retired electric vehicle batteries, which are referred to as EV batteries throughout the paper. The proposed method is constructed around surface data, which include battery capacity, service time, cycle number, appearance damage, repair history, battery health, market prices of critical metals, battery retirement volume, and price indices. These data are convenient to obtain on a recycling site without opening or testing the internal cells. By combining a novel bidirectional denoising autoencoder (BDAE) with a support vector regression (SVR) model optimized by the grey wolf optimizer (GWO), we are able to extract meaningful representations from noisy surface data and map those representations to recycling prices. The overall approach is motivated by the idea that noisy and redundant input factors may hide the true value-related signal. Therefore, before making a prediction, the data should be transformed into a low-dimensional feature space where the influence of noise is suppressed and the main patterns relevant to price are retained.
The figure above conceptually illustrates a retired EV battery module awaiting value assessment. In practice, many such modules enter the recycling chain every day. The recycling process normally starts with logistics collection and is followed by inspection, disassembly, and further utilization. If the recycling price can be estimated accurately on-site using surface data, then unnecessary transportation and dismantling costs can be avoided. In particular, the possibility of echelon utilization can be preliminary evaluated before deciding whether to spend resources on a deep inspection. Thus, our research goal is not simply to refine an algorithm but to provide a practical, data-driven tool to support EV battery recycling enterprises in making quick and profitable decisions.
1. Characterization of the EV Battery Recycling Price Prediction Problem
1.1 Surface influencing factors for retired EV batteries
Many possible factors can influence the price of a retired EV battery. We split them into two broad categories: the intrinsic state of the physical battery and the external market context. In terms of the intrinsic state, the surface-level factors are those that can be observed or logged during battery usage. Among them, initial parameters such as battery capacity directly reflect the potential storage amount. The personalized parameters include battery service time, charge-discharge cycle count, appearance abrasion level, repair/maintenance count, and state of health. These parameters are often recorded by the battery management system or by external inspection. A battery that has been used for a shorter period, has fewer cycles, and has received less damage usually retains higher residual value and may be reused for echelon applications. On the other hand, batteries with a large number of repairs, severe deformation, or low health status may only be suitable for material recycling, and their recycling price should be lower.
From the market dimension, the most relevant data include spot prices of nickel, cobalt, manganese, and lithium, because these metals are the main value carriers in the cathode active material. When these metal prices rise, the recycling revenue for the corresponding metals increases, and recyclers can afford to pay more for retired EV batteries. The number of retired EV batteries in a certain period also represents supply, which strongly affects market equilibrium. Moreover, the price index of new power batteries and the price index of retired power batteries reflect the overall momentum of the battery market. We therefore use those 13 surface factors, which are summarized in Table 1, as the initial inputs to our prediction system.
| Category | Factor name | Symbol |
|---|---|---|
| Battery intrinsic factors | Battery capacity | \(x_1\) |
| Service time | \(x_2\) | |
| Battery individual factors | Cycle number | \(x_3\) |
| Appearance wear degree | \(x_4\) | |
| Repair times | \(x_5\) | |
| Battery state of health | \(x_6\) | |
| Market factors | Nickel market price | \(x_7\) |
| Cobalt market price | \(x_8\) | |
| Manganese market price | \(x_9\) | |
| Lithium market price | \(x_{10}\) | |
| Retired battery quantity | \(x_{11}\) | |
| New power battery price index | \(x_{12}\) | |
| Retired EV battery price index | \(x_{13}\) |
1.2 Correlation analysis of surface factors
After selecting the candidate factors, we need to determine whether these factors are statistically related to the recycling price. We applied Spearman’s rank correlation coefficient because the relationships could be nonlinear and because rank-based correlation is less sensitive to outliers. The coefficient between a surface factor ranking and the recycling price ranking is defined as
\[
r_s = 1-\frac{6\sum_{i=1}^n (x_i-y_i)^2}{n(n^2-1)}
\]
where \(x_i\) is the rank of the \(i\)-th value for one surface factor, \(y_i\) is the rank of the \(i\)-th recycling price, and \(n\) is the number of retired EV battery samples. A value approaching 1 or -1 indicates a strong positive or negative monotonic relation, whereas a value near 0 indicates almost no correlation. In the experimental dataset obtained from a battery recycling enterprise, the Spearman coefficients of all selected 13 factors exceeded 0.1 and, more importantly, were statistically meaningful in cross-validation. Factors with extremely low correlation were removed so that the remaining factors all contributed to the price signal to some extent.
1.3 Challenges in surface-data-driven prediction
The major difficulty of using surface data to predict EV battery recycling price is that surface factors are redundant and mutually correlated. For example, the service time of a retired EV battery is strongly correlated with the number of cycles and also with the appearance wear. The price indices of new and retired batteries also move together. This multicollinearity violates the assumptions of many linear models and causes unstable coefficient estimates. Furthermore, data collected from recycling sites often contain measurement errors or missing values, adding noise to the explanatory variables. The intrinsic relationship between surface data and recycling price is weak in the raw feature space, so a prediction model may fit the noise instead of the underlying pattern. To overcome these challenges, we first perform denoising and feature extraction to create compact and informative features, and then apply a robust regression model whose hyperparameters are automatically optimized.
1.4 Overall prediction framework
Our framework contains three major stages. In the first stage, Spearman correlation analysis is used to select relevant surface factors. In the second stage, a bidirectional denoising autoencoder is constructed to learn low-dimensional representations from the noisy factors. The autoencoder takes the 13 selected factors as input, applies random noise injection during training, and learns to reconstruct the denoised data. In the third stage, the extracted feature vector is fed into a SVR model. The GWO is utilized to search for the optimal penalty coefficient \(C\) and kernel width \(\sigma\) of SVR automatically. This combined architecture is abbreviated as BDAE-GWO-SVR.
2. Surface Data Preprocessing and Feature Extraction
2.1 Preprocessing of EV battery surface data
The raw surface data contain duplicate records, wrong price entries, missing values, and incomparable scales. Therefore, a data cleaning step is first applied. Duplicate records are removed. Missing capacity values are replaced by the median capacity of the same battery type. Incorrect price values that are far outside the normal range are treated as outliers and eliminated. After cleaning, min-max normalization is used to scale all features to the range \([0,1]\),
\[
x’=\frac{x-\min(x)}{\max(x)-\min(x)}
\]
where \(x’\) is the normalized value of an input feature. This procedure preserves the relative magnitude and distribution of the original EV battery data, while ensuring that the model training process converges smoothly.
2.2 Autoencoder and denoising mechanism
An autoencoder is an unsupervised neural network that learns to reconstruct its own input through a bottleneck hidden layer. Let \(\mathbf{X}=[x_1,\dots,x_{13}]\) denote the normalized surface factors. The encoder maps the input to a hidden representation \(\mathbf{H}_t\), and the decoder attempts to reconstruct the original input from this hidden representation. A plain autoencoder minimizes a reconstruction error such as mean square error. However, because the input data contain noise, we adopt a denoising autoencoder. In a denoising autoencoder, artificially corrupted versions \(\tilde{\mathbf{X}}\) of the input are generated by adding random noise. The network is trained to reconstruct the clean data from the corrupted input. This forces the encoder to discover robust features that are insensitive to noise. The encoding process is
\[
\mathbf{H}_t = s_f(\mathbf{W}\tilde{\mathbf{X}}+\mathbf{b})
\]
where \(\mathbf{W}\) and \(\mathbf{b}\) are the weights and biases of the encoding layer, and \(s_f(\cdot)\) is the activation function. The decoding process is
\[
\mathbf{Y}_t = s_g(\mathbf{W}’\mathbf{H}_t+\mathbf{b}’)
\]
where \(\mathbf{W}’\) and \(\mathbf{b}’\) are decoder weight and bias values and \(s_g(\cdot)\) is the activation function. The reconstruction error is
\[
L = \frac{1}{N}\sum_{i=1}^{N}\lVert \mathbf{x}^{(i)}-\hat{\mathbf{x}}^{(i)}\rVert^2
\]
where \(\mathbf{x}^{(i)}\) is the clean input of the \(i\)-th EV battery sample, \(\hat{\mathbf{x}}^{(i)}\) is its reconstruction, and \(N\) is the sample size.
2.3 Bidirectional denoising autoencoder architecture
Although a single denoising autoencoder can suppress random noise, surface factors have complex bidirectional dependencies. For example, the recycling price also has a backward influence on market behavior because high recovery prices may incentivize more battery owners to turn in their old batteries. To capture such two-way dependencies, we designed a bidirectional denoising autoencoder (BDAE). The BDAE receives both the forward surface data \(\vec{\mathbf{X}}\) and the reverse surface data \(\overleftarrow{\mathbf{X}}\). The forward and reverse streams share similar encoder-decoder structures but have independent weights. Their forward hidden representation is
\[
\overrightarrow{\mathbf{H}}_t = \sigma(\vec{\mathbf{W}}_1\vec{\mathbf{X}}+\vec{\mathbf{b}}_1)
\]
and the reverse hidden representation is
\[
\overleftarrow{\mathbf{H}}_t = \sigma(\overleftarrow{\mathbf{W}}_1\overleftarrow{\mathbf{X}}+\overleftarrow{\mathbf{b}}_1)
\]
The final feature representation is formed by merging the forward and reverse hidden layers according to a weighted sum:
\[
\mathbf{H}_t = w_1\overrightarrow{\mathbf{H}}_t + w_2\overleftarrow{\mathbf{H}}_t
\]
The decoder reconstructs the denoised data from \(\mathbf{H}_t\),
\[
\mathbf{Y}_t = \sigma(\mathbf{W}_2\mathbf{H}_t+\mathbf{b}_2)
\]
where \(w_1\) and \(w_2\) are learnable weights balancing the two directions. Because the input is the same set of 13 EV battery factors in both directions, the BDAE does not increase the required amount of data. Instead, it improves feature extraction by compensating for the bias of a one-way model and exploiting complementary information from the reversed feature order. The output of the hidden layer \(\mathbf{H}_t\) is the extracted deep feature vector that is used as the input of the price prediction model.
2.4 Parameter settings for the BDAE
The BDAE was designed with an input layer of 13 nodes, two hidden layers, and an output feature layer of 6 nodes. The first hidden layer had 10 nodes, and the second hidden layer had 8 nodes. We selected the ELU activation function because of its smooth nonlinear behavior and its ability to mitigate the vanishing gradient problem in deep networks. The training configuration is listed in Table 2.
| Parameter | Value |
|---|---|
| Input layer nodes | 13 |
| First hidden layer nodes | 10 |
| Second hidden layer nodes | 8 |
| Output layer nodes | 6 |
| Batch size | 32 |
| Epochs | 2000 |
| Learning rate | 0.0001 |
| Noise type | Gaussian |
During training, the loss function \(L\) is minimized by backpropagation. The weights and biases are updated according to
\[
W_{t+1}=W_t-\alpha\frac{\partial L}{\partial W_t}
\]
\[
b_{t+1}=b_t-\alpha\frac{\partial L}{\partial b_t}
\]
where \(\alpha\) is the learning rate and \(t\) denotes the iteration number. The training continues until the reconstruction loss converges. The feature extraction process is described in Table 3.
| Step | Operation |
|---|---|
| Step 1 | Initialize the learning rate, weights, and biases of the BDAE. |
| Step 2 | Take the 13 surface factors of retired EV batteries as input, and compute the forward and reverse hidden representations. |
| Step 3 | Merge the forward and reverse hidden layers to obtain final low-dimensional features. |
| Step 4 | Reconstruct the denoised input through the decoder. |
| Step 5 | Compute the reconstruction error between the original and reconstructed data. |
| Step 6 | Update all parameters using gradient descent. |
| Step 7 | Repeat until the loss function converges and output the extracted features. |
3. Surface Feature Factor Mining Based on BDAE
3.1 Feature extraction performance on a benchmark example
To evaluate the capability of the BDAE feature extraction module, we first used a widely known regression dataset whose input attributes resemble the categorical and continuous nature of EV battery surface data. The dataset contains many quantitative features such as lot area, overall quality, year built, living area, and number of bathrooms. We selected 14 features from that dataset and normalized all of them. We split the dataset into four ratios, namely 6:4, 7:3, 8:2, and 9:1. The reconstruction errors for the EV battery-oriented feature extraction model were measured using the same architecture as described before but with 14 input features. The result indicated that the BDAE reconstruction error remained around 0.005 under all splits, which is a very low level. This result suggests that the BDAE can accurately reconstruct the input while compressing the feature representation, and thus the retained hidden features are informative.
3.2 Comparison with other feature extraction models
In addition, we compared the BDAE with a deep belief network (DBN) and a convolutional neural network (CNN) on the same benchmark data. Using a 7:3 training-test split, the BDAE achieved a reconstruction error of 0.0048. In contrast, DBN produced 0.018 and CNN produced 0.023. Table 4 lists the comparison results. The significantly lower error of BDAE is attributed to its bidirectional denoising mechanism, which reduces the influence of random noise and better captures the internal correlations of the attributes. The strong performance on this benchmark provides confidence that BDAE can effectively extract quality features from the noisy surface data of retired EV batteries.
| Model | Training : Test Split | Reconstruction Loss |
|---|---|---|
| BDAE | 6:4 | 0.0051 |
| BDAE | 7:3 | 0.0048 |
| BDAE | 8:2 | 0.0045 |
| BDAE | 9:1 | 0.0050 |
| DBN | 7:3 | 0.018 |
| CNN | 7:3 | 0.023 |
4. GWO-Optimized SVR for EV Battery Recycling Price Prediction
4.1 Support vector regression for EV battery recycling price
Support vector regression is an effective tool for nonlinear regression, especially when the number of training samples is limited and the input features have been properly preprocessed. Given the extracted feature vector \(\mathbf{H}_t=[h_1,\dots,h_6]\), the SVR model seeks an optimal mapping from \(\mathbf{H}_t\) to the recycling price \(y\). The nonlinear regression function is
\[
y=W\cdot\phi(\mathbf{H}_t)+b
\]
where \(\phi(\cdot)\) maps the features into a higher-dimensional space, \(W\) is the regression weight vector, and \(b\) is the intercept. The primal optimization problem can be written as
\[
\min_{W,b}\frac{1}{2}\lVert W\rVert^2+C\sum_{i=1}^{l}(\xi_i+\xi_i^*)
\]
subject to the constraints that all points lie within an \(\varepsilon\)-insensitive tube, except for the slack variables \(\xi_i,\xi_i^*>0\). The dual optimization problem incorporates a kernel function \(K(\cdot,\cdot)\),
\[
\max\left[-\frac{1}{2}\sum_{i=1}^{m}\sum_{j=1}^{m}(a_i-a_i^*)(a_j-a_j^*)K(\mathbf{h}_i,\mathbf{h}_j)-\sum_{i=1}^{m}a_i^*(y_i+\varepsilon)+\sum_{i=1}^{m}a_i(y_i-\varepsilon)\right]
\]
where \(a_i,a_i^*\) are Lagrange multipliers. We chose the radial basis function (RBF) kernel,
\[
K(\mathbf{h}_i,\mathbf{h}_j)=\exp\left(-\frac{\lVert\mathbf{h}_i-\mathbf{h}_j\rVert^2}{2\sigma^2}\right)
\]
The RBF kernel can approximate complex nonlinear relationships between the surface features and the EV battery recycling price. The resulting regression function is
\[
\hat{y}=\sum_{i\in SV}(a_i-a_i^*)K(\mathbf{h}_i,\mathbf{h})+b
\]
Because the number of support vectors is generally small, SVR does not require a huge amount of data to avoid overfitting, which is an important advantage in the field of retired EV battery recycling where data collection is still growing.
4.2 Grey wolf optimizer for hyperparameter search
In SVR, the regularization parameter \(C\) and the RBF kernel width \(\sigma\) strongly affect both training accuracy and generalization. In our EV battery recycling data, the variation of different battery models and usage patterns is large. If we choose \(C\) and \(\sigma\) by a simple grid search or by experience, the model may work well for one battery type but poorly for another. Therefore, we used the grey wolf optimizer to perform an automatic search for the optimal hyperparameter pair.
GWO mimics the leadership hierarchy and hunting behavior of grey wolves. In GWO, a population of wolves represents candidate pairs \((C,\sigma)\). The best solution is denoted as \(\alpha\), the second best as \(\beta\), and the third best as \(\delta\). All other wolves are classified as \(\omega\). The hunting process is iterated until a maximum number of generations. The linear convergence factor \(a\) is updated by
\[
a=2-\frac{2t}{T_{\max}}
\]
where \(t\) is the current iteration and \(T_{\max}\) is the maximum number of iterations. The coefficient vectors are
\[
A=2a\cdot r_2-a
\]
\[
C=2r_1
\]
where \(r_1\) and \(r_2\) are random vectors uniformly distributed in \([0,1]\). For each wolf, the encircling movement is computed according to the three leading wolves,
\[
D_{\alpha}=|C_1 X_{\alpha}-X(t)|
\]
\[
X_1=X_{\alpha}-A_1D_{\alpha}
\]
and analogous expressions are used for \(\beta\) and \(\delta\). The updated position is obtained by
\[
X(t+1)=\frac{X_1+X_2+X_3}{3}
\]
The fitness of each candidate pair is evaluated using ten-fold cross-validation on the training dataset. The accuracy of each SVR model is measured by the coefficient of determination of the cross-validation, \(R^2_{cv}\). Let \(RSS_j\) and \(TSS_j\) be the residual sum of squares and the total sum of squares for the \(j\)-th fold. Then,
\[
R^2_{cv}=1-\frac{\sum_{j=1}^{k}RSS_j}{\sum_{j=1}^{k}TSS_j}
\]
where \(k=10\). A larger \(R^2_{cv}\) corresponds to better predictive performance. GWO searches for the \((C,\sigma)\) pair that maximizes \(R^2_{cv}\).
4.3 Evaluation metrics
We use two standard metrics to assess the final prediction accuracy: root mean square error (RMSE) and mean absolute percentage error (MAPE). They are defined as
\[
RMSE=\sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i-\hat{y}_i)^2}
\]
\[
MAPE=\frac{100\%}{n}\sum_{i=1}^{n}\left|\frac{y_i-\hat{y}_i}{y_i}\right|
\]
where \(y_i\) is the actual recycling price and \(\hat{y}_i\) is the predicted price. RMSE measures the absolute error level, while MAPE measures the relative error. Both are smaller for better prediction accuracy.
5. Development of an EV Battery Recycling Price Prediction Support System
5.1 Case data description
To validate the whole method, we applied the model to a dataset containing 5000 records of retired EV batteries collected from a recycling enterprise. These records cover five different models from the same specification family. Each record contains the 13 surface factors and the corresponding actual recycling price. The data include battery capacity from 65 to 102 kWh, service time from about 34,000 to 45,000 hours, and cycle numbers ranging from about 5,000 to 10,000. The market prices of nickel, cobalt, manganese, and lithium were collected from an official trading platform. The actual recycling price values varied from around 860 to 1350 dollars. A sample of the data is shown in Table 5.
| Sample ID | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| Capacity (kWh) | 100 | 85 | 102 | 95 | 70 | 80 | 77 | 65 |
| Service time (h) | 34566 | 45473 | 42387 | 42898 | 38443 | 36283 | 37829 | 42781 |
| Cycle number | 7876 | 7865 | 6534 | 8723 | 8532 | 9876 | 5437 | 6234 |
| Repair times | 3 | 2 | 4 | 1 | 5 | 0 | 0 | 3 |
| State of health (%) | 81 | 80 | 82 | 75 | 77 | 68 | 85 | 77 |
| Nickel price ($/kg) | 36.25 | 43.75 | 45.97 | 37.36 | 44.72 | 44.86 | 45.00 | 43.05 |
| Cobalt price ($/kg) | 36.94 | 36.80 | 36.52 | 37.36 | 37.22 | 37.08 | 38.09 | 36.67 |
| Manganese price ($/kg) | 18.33 | 18.88 | 18.87 | 17.92 | 17.90 | 18.19 | 17.78 | 18.89 |
| Lithium price ($/kg) | 92.36 | 91.94 | 92.91 | 90.69 | 91.38 | 92.98 | 90.13 | 90.83 |
| Retired battery quantity | 2654 | 3564 | 3632 | 2964 | 3463 | 2659 | 3459 | 6543 |
| New battery price index (%) | 1.1 | -2.2 | -1.8 | 0.3 | 0.9 | 2.3 | -3.6 | -0.2 |
| Retired battery price index (%) | -2.2 | -3.1 | -0.4 | 0.9 | 3.1 | 1.3 | -1.4 | 1.2 |
| Appearance wear degree | 5 | 1 | 1 | 1 | 5 | 3 | 1 | 2 |
| Recycling price ($) | 1354.56 | 898.86 | 1006.92 | 907.91 | 863.89 | 937.68 | 1120.21 | 1243.91 |
5.2 Software system architecture
We developed a web-based decision support system for EV battery recycling price prediction using Python and PyQt5. The system adopts a browser/server architecture, consisting of a user layer, an application layer, and a data layer. In the user layer, recycling staff can enter battery information, upload data files, and view prediction results. The application layer integrates the BDAE feature extraction model and the GWO-SVR price prediction model. It also provides model training, parameter optimization, and prediction functions. The data layer stores the historical EV battery information database, market price database, and past prediction records. Table 6 summarizes the development environment.
| Type | Configuration |
|---|---|
| Operating System | Windows 10 |
| Programming Language | Python |
| GUI Framework | PyQt5 |
| GPU | NVIDIA GeForce GTX 4060 |
| CPU | AMD R9-7945HX (2.50 GHz) |
| Memory | 16 GB |
5.3 Main functional modules
The proposed support system includes four major functional modules: user login and management, data center, configuration center, and control panel. The login module authenticates users and provides password recovery and account registration. The data center displays the intrinsic battery factors and market factors. Users can add, delete, and query historical data. The configuration center is the core parameter setting interface for the BDAE and SVR models. The BDAE parameters include the number of input nodes, number of hidden nodes, batch size, epochs, and learning rate. The GWO parameters include the upper and lower boundaries of the search space and the maximum number of iterations. The control panel allows users to enter a new EV battery’s capacity, service time, cycle count, and health state. After clicking the prediction button, the system calculates the estimated recycling price and displays it on the same page.
5.4 Experimental results and comparison
We trained the BDAE-GWO-SVR model on 80% of the data and tested it on the remaining 20%. The GWO search was allowed to run for 20 iterations. During the optimization, the accuracy of the SVR model increased rapidly in the first two iterations and then remained stable. The final cross-validation accuracy exceeded 0.81. The optimized hyperparameters enabled the final SVR to produce good predictions on the test set. The overall RMSE and MAPE of the proposed model were 0.371 thousand dollars and 0.060, respectively.
To understand the contribution of each component, we compared the proposed BDAE-GWO-SVR model with three alternative models: plain SVR, GWO-SVR without BDAE, and two additional baseline models, namely random forest (RF) and back-propagation neural network (BP). The prediction errors are summarized in Table 7.
| Model | RMSE (thousand dollars) | MAPE |
|---|---|---|
| SVR | 0.453 | 0.073 |
| GWO-SVR | 0.413 | 0.067 |
| BDAE-GWO-SVR | 0.371 | 0.060 |
| Random Forest (RF) | 1.058 | 0.084 |
| Back-propagation Neural Network (BP) | 0.812 | 0.080 |
From Table 7, we observe that the BDAE-GWO-SVR model achieves the smallest RMSE and MAPE among all compared methods. The implementation of GWO improves the prediction accuracy compared with the standard SVR because it finds a better hyperparameter combination for the given EV battery data. The addition of BDAE further reduces both RMSE and MAPE. This suggests that the denoising feature extraction step successfully removes redundancy and noise from the surface factors and provides a cleaner feature set for the SVR model. The RF and BP models achieve inferior results, which illustrates that neural networks may not be suitable when the sample size is moderate and the noise level is high.
We also conducted an ablation study to visualize the prediction points against the actual recycling prices. The predicted values obtained by the direct SVR generally followed the trend of the actual prices, but some deviations appeared near peak prices. After the BDAE-GWO denoising procedure, the predicted points were much closer to the real values. This behavior reinforces the practical value of the proposed method for EV battery recycling price estimation.
6. Discussion
6.1 Interpretation of the results
Our experimental study has demonstrated two key points. First, surface data alone can provide enough information for a useful prediction of the recycling price if they are processed through a denoising feature extraction method. The bidirectional structure of BDAE is beneficial because it permits the model to capture both forward and backward relationships among the factors. For instance, a low battery health value may decrease the price, but a high cobalt market price may compensate for that decline. By learning these mutual dependencies, BDAE compresses the 13 raw factors into six representational features that retain the most relevant price-oriented information.
Second, the choice of the regression model is also critical. In our dataset, neural networks and random forests failed to reach the same accuracy as the SVR model. The small number of labeled EV battery samples and the noisy nature of the data are typical in real recycling enterprises. SVR is known for its strong generalization in such settings. By optimizing the two hyperparameters with GWO, the model becomes more adaptive to the particular distribution of data from different EV battery models. Therefore, the combination of BDAE and GWO-SVR is superior to using either of them independently.
6.2 Limitations
One limitation of the proposed system is that the surface data still cannot fully reflect the internal residual value of a retired EV battery. For example, the true equivalent internal resistance and the remaining capacity under high current cannot be measured without charging-discharging tests. Another limitation is that our model is trained on historical data from a limited set of battery models. If a brand-new battery chemistry or a completely different recycling market emerges, the model should be retrained or updated. In addition, the current model does not incorporate textual information such as battery brand nuance or environmental policy changes. These factors could be considered in future work by adding natural language processing and dynamic feature selection.
7. Conclusion and Future Work
In this article, we have presented a complete data-driven framework for predicting the recycling price of retired EV batteries using only surface-level influencing factors. The method starts with 13 selected factors from both the battery intrinsic state and the market environment. A bidirectional denoising autoencoder, BDAE, is constructed to denoise the raw data and extract compact features whose correlation with the recycling price is strengthened. The extracted features are then used as input to a support vector regression model whose hyperparameters are automatically optimized by the grey wolf optimizer. The proposed BDAE-GWO-SVR method was evaluated on real retired EV battery records. The experimental results show that it outperforms plain SVR, GWO-SVR, random forest, and back-propagation neural networks, achieving a MAPE of only 6%. The developed web-based support system allows recycling enterprises to enter battery surface data and quickly obtain an estimated recycling price, which can support decisions about transport, disassembly, and echelon utilization.
Future research can proceed in several directions. First, additional external information such as policy intensity, subsidy levels, and carbon trading prices should be integrated into the input factors because these elements increasingly affect the market value of retired EV batteries. Second, deep leaning attention mechanisms could be combined with BDAE to help identify which features are most important for different battery categories. Third, interpretability techniques such as SHapley Additive exPlanations should be introduced to explain why a certain recycling price is predicted for a particular EV battery. Finally, the support system can be improved with more interactive visualization and real-time data updating capabilities, making it easier for end users to monitor price trends and recalibrate the model as new data become available.